Optimizing the Retrieval of Pertinent Answers for NL Questions with the E-Librarian Service 1

Abstract

Abstract. There is a growing discrepancy between the creation of digital content and its actual employment and usefulness in a learning society. Technologies for recording lectures have become readily available and the sheer number and size of such objects produced grows exponentially. However, in practice most recordings are monolithic entities that cannot be integrated into an active learning process offhand. To overcome this problem, recorded lectures have to be semantically annotated to become full-fledged e-learning objects facilitating automated reasoning over their content. We present a running web-based system — the e-Librarian Service CHESt — that is able to match a user’s question given in natural language to a selection of semantically pertinent learning objects based on an adapted best cover algorithm. We show with empirical data that the precision of our e-Librarian Service is much more efficient than traditional keyword-based information retrieval; it yields a correct answer in most of the cases (93 % of the queries), and mostly with a high precision, i.e., without supplementary hits. We also describe some ideas to improve the retrieval performance by user feedback. 1